Neural Polarization: Toward Electron Density for Molecules by Extending Equivariant Networks
This work addresses the problem of accurately modeling molecular properties for computational chemistry by extending equivariant networks, though it is incremental as it builds on existing methods.
The authors tackled the limitation of existing SO(3)-equivariant models in representing electron density and polarization effects in molecules by proposing Neural Polarization, which embeds atoms as fixed and moving points to include electron density, resulting in improved prediction performances across various targets.
Recent SO(3)-equivariant models embedded a molecule as a set of single atoms fixed in the three-dimensional space, which is analogous to a ball-and-stick view. This perspective provides a concise view of atom arrangements, however, the surrounding electron density cannot be represented and its polarization effects may be underestimated. To overcome this limitation, we propose \textit{Neural Polarization}, a novel method extending equivariant network by embedding each atom as a pair of fixed and moving points. Motivated by density functional theory, Neural Polarization represents molecules as a space-filling view which includes an electron density, in contrast with a ball-and-stick view. Neural Polarization can flexibly be applied to most type of existing equivariant models. We showed that Neural Polarization can improve prediction performances of existing models over a wide range of targets. Finally, we verified that our method can improve the expressiveness and equivariance in terms of mathematical aspects.